🧮 Ethosoft-Math-Translator

English → Turkish translation for mathematics — formulas, symbols and numbers survive intact.

General-purpose MT models are good at prose and terrible at math: they translate \frac{a}{b} into nonsense, "fix" decimal points, drop ±, √, ∑, CJK characters, and silently change numbers. Ethosoft-Math-Translator is NedoTranslator fine-tuned so that mathematical content is treated as untouchable: it translates the words around the math and hands every formula back byte-for-byte.

EN  Since $x^2-4=(x-2)(x+2)$, the solutions are x = ±2, so the sum is 0.
TR  $x^2-4=(x-2)(x+2)$ olduğundan, çözümler x = ±2'dir, bu yüzden toplam 0'dır.

EN  Compute $\sqrt{16} + 3^2$. First, $\sqrt{16}=4$ and $3^2=9$, hence the total is $13$.
TR  $\sqrt{16} + 3^2$ değerini hesaplayın. Önce $\sqrt{16}=4$ ve $3^2=9$, dolayısıyla toplam $13$'tür.

EN  Evaluate ∑_{k=1}^{10} k × 2 ≤ 110. 你好
TR  ∑_{k=1}^{10} k × 2 ≤ 110 değerini hesaplayın. 你好

(real, unedited outputs of the released checkpoint)

How it works

  text ──► mask ──► translate line by line ──► check ──► restore ──► Turkish text
           │                                     │
           │  LaTeX ($..$, $$..$$, \[..\], \begin..\end),
           │  code, URLs, tool-call blocks, math Unicode (±×÷√∑∫≤≥≠…),
           │  CJK, and characters missing from the vocabulary
           └─► become  [[1]] [[2]] …  (renumbered per line)

The model was fine-tuned to copy [[n]] placeholders exactly while translating everything else. After translation every placeholder is verified and substituted back; a line whose placeholders were damaged is rejected instead of being silently corrupted. This is why formulas can never be "translated" or altered.

Quick start

The easiest way is the Ethosoft-Math library, which ships the masking / restoring pipeline:

pip install "Ethosoft-Math[translate]"

from ethosoft_math.translate import MathTranslator

t = MathTranslator()   # downloads Ethosoft/Ethosoft-Math-Translator
print(t.translate("If $x^2 = 9$, then x = ±3.\nThe answer is $\\boxed{3}$ or $-3$."))
# Eğer $x^2 = 9$ ise, x = ±3.
# Cevap $\boxed{3}$ veya $-3$'tür.

# whole records (e.g. dataset rows); returns None if any formula could not be preserved
row = t.translate_record({"problem": "What is $\\frac{1}{2}+\\frac{1}{3}$?", "solution": "It is $\\frac{5}{6}$."})

Or from the command line:

ethosoft-math translate "Compute $\\sqrt{16} + 3^2$."
ethosoft-math translate --input data.jsonl --output data_tr.jsonl     # adds problem_tr / solution_tr

Plain 🤗 Transformers

The raw model works on already masked text. If you call it directly, mask math yourself (see ethosoft_math.translate.protect_all) – feeding raw LaTeX to the model is not what it was trained for.

from transformers import MarianMTModel, MarianTokenizer
tok = MarianTokenizer.from_pretrained("Ethosoft/Ethosoft-Math-Translator")
model = MarianMTModel.from_pretrained("Ethosoft/Ethosoft-Math-Translator")

batch = tok(["Therefore the value of [[1]] is [[2]]."], return_tensors="pt")
out = model.generate(**batch, num_beams=4, max_new_tokens=256)
print(tok.batch_decode(out, skip_special_tokens=True))   # Turkish sentence with [[1]] and [[2]] kept verbatim

Results

Held-out evaluation (rows never seen in fine-tuning; hash-based split):

Metric	NedoTranslator (base)	Ethosoft-Math-Translator
Validation loss (masked lines)	0.701	0.373
Placeholder integrity, line level (all [[n]] reproduced)	68.5 %	99.3 %
End-to-end rows passing all quality gates¹ (276 unseen rows)	–	94.9 %

¹ numbers unchanged, formulas unchanged, answer preserved, no foreign scripts, Turkish output, same structure. Of the 5.1 % rejected rows, 11 were placeholder failures and 3 changed a number – they are dropped, not emitted.

On sampled rows the output is close to the teacher's translations (see below); we did not run a human evaluation, so treat fluency claims accordingly.

In production: 755,035 rows translated

The released checkpoint was run over the full pool of math corpora (on top of the 27,768 teacher-translated rows used for training). Every row went through the same quality gate; rows that failed were deleted, not repaired.

Dataset	Translated	Rejected	Pass rate
MetaMathQA	382,532	8,180	97.9 %
MathInstruct	111,186	6,061	94.8 %
OpenR1-Math	78,894	10,767	88.0 %
NuminaMath-TIR	64,772	4,944	92.9 %
Orca-Math	59,239	2,584	95.8 %
AQuA-RAT	40,103	1,295	96.9 %
MATH	7,886	858	90.2 %
GSM8K	6,228	171	97.3 %
DeepScaleR	4,195	1,537	73.2 %
Total	755,035	36,397	95.4 %

Main rejection reasons: a placeholder was damaged (17.7 k), a number changed (17.2 k), answer lost (0.6 k), formula changed (0.5 k), output not Turkish (0.4 k). Throughput on a single H100: ~10–19 rows/s (beam 4). Hard competition sets such as DeepScaleR lose the most rows because their solutions are dense with formulas.

Training

• Base: nmsofficial/NedoTranslator (MarianMT, 236.9 M parameters), itself fine-tuned from Helsinki-NLP/opus-mt-tc-big-en-tr.
• Teacher data: 27,768 math problems + solutions translated into Turkish by a strong LLM (DeepSeek-V4 Flash, reasoning off, via the Evren LLM gateway) under a strict quality gate, then masked and split into 161,085 aligned line pairs (1,412 held out). Source corpora: GSM8K, MATH, MetaMathQA, Orca-Math, MathInstruct, OpenR1-Math, NuminaMath-TIR, DeepScaleR, AQuA-RAT.
• Quality gate for teacher output (rows failing any check were deleted, ~4–10 % per dataset): number multiset preserved · every formula block preserved · gold answer still present · every CJK/math symbol of the source present · no foreign script added · Turkish-ness · line structure.
• Recipe: 3 epochs, AdamW, lr 5e-5 (linear warm-up/decay), label smoothing 0.1, bf16, ~12 k tokens/batch, 1 × H100 (≈ 4 minutes), TRUBA HPC.

Intended use & limitations

• ✅ Translating math word problems, solutions, proofs and textbook-style exposition EN → TR, especially to build Turkish math datasets.
• ✅ Pipelines that must keep formulas/numbers exactly (the library rejects rows it cannot guarantee).
• ⚠️ Text inside masked blocks (e.g. \text{if } n … inside a LaTeX block, or code comments) is not translated.
• ⚠️ Translations inherit the teacher model's style and mistakes; terminology is not hand-verified.
• ⚠️ Long lines are truncated at 256 tokens – the library splits input by line, so keep one idea per line.
• ⚠️ English → Turkish only.

License

Released under CC BY-NC 4.0. The upstream NedoTranslator card lists cc-by-nc-2.0 in its metadata and "CC BY 4.0" in its text; we adopt the more restrictive non-commercial terms until that is clarified. Please also retain attribution to the upstream models below.

Attribution & citation

This work stands on NedoTranslator by nmsofficial — please credit it: https://huggingface.co/nmsofficial/NedoTranslator — which in turn builds on Helsinki-NLP's OPUS-MT:

@inproceedings{tiedemann-thottingal-2020-opus,
  title     = {{OPUS-MT} -- Building open translation services for the World},
  author    = {Tiedemann, J{\"o}rg and Thottingal, Santhosh},
  booktitle = {Proceedings of the 22nd Annual Conference of the European Association for Machine Translation},
  year      = {2020}
}

If you use Ethosoft-Math-Translator:

@misc{ethosoft_math_translator_2026,
  title        = {Ethosoft-Math-Translator: Symbol-Preserving English-to-Turkish Mathematical Translation},
  author       = {Ethosoft},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/Ethosoft/Ethosoft-Math-Translator}},
  note         = {Fine-tuned from NedoTranslator (nmsofficial)}
}

_{Made by Ethosoft · Ethosoft-Math library: github.com/ethosoftai/ethosoft-math}